Dual Channel Consistent Embedding Prototypical Networks for Industrial Few-Shot Fault Diagnosis

Fault diagnosis plays a crucial role in industrial systems and various supervised learning approaches have been extensively employed in this domain. These methods typically require substantial amounts of labeled fault data for effective model training. However, the available faulty samples in real industrial processes are limited. In this study, a dual-channel consistent embedding prototypical networks (DCEPN) is proposed to improve the performance of few-shot fault diagnosis. The DCEPN includes the prototypical channel and the semantic channel, which the former is metric-based prototypical networks (PN) and the latter incorporates detailed fault descriptions to improve the discriminative ability of prototypes learned. In the semantic channel, fault descriptions are artificially transformed into binary-value attribution vectors and further are mapped into semantic features to predict the real fault types. To align prototypes and semantic features in dual embedding space, an effective metric matrix is designed to evaluate the consistency, where diagonal elements correspond to intra-dimension consistency, while off-diagonal elements capture cross-dimensional consistency relationships. Comparative experiments with four state-of-the-art few-shot learning approaches validate the effectiveness and superior performance of the proposed DCEPN.